Senior Dev: This "Grill Me" Prompt Is Going Viral Among Top Engineers

From Total TypeScript to AI Hero

At the AI Engineer Europe booth in London, Matt Pocock shows up in full teacher mode: joking about the event, then immediately grounding the conversation in what he actually does. He explains why he now runs two-week cohort courses instead of classic self-paced products — AI changes too fast, and he’s already been burned by timing, from AI SDK v5 dropping right after a v4-based course to Claude updates landing mid-course.

Why “code doesn’t matter anymore” keeps failing in practice

Pocock’s big thesis is a rebuttal to the popular idea that code is now just a compile target for English. Every time he tried to ignore the code and let the model do the rest, he says, he ended up with a terrible mess — which pushed him back to classic engineering texts like The Pragmatic Programmer, extreme programming, Philosophy of Software Design, and DDD. His takeaway is simple: if a codebase is easy for humans to change, it’s easier for AI to change too; if it’s painful for humans, it’s even worse for models.

Deep modules, narrow waists, and containing the slop

The hosts riff on “narrow waist” architecture — define what goes in and out clearly, and let the internals vary — and Pocock connects that to John Ousterhout’s “deep modules.” The vibe here is very practical: humans should own the interfaces and architecture, then delegate the inside of the box to AI. That’s where he says AI coding works best: not as magic, but as leverage inside intentionally designed boundaries.

Why DDD suddenly feels made for prompting

Pocock lights up talking about domain-driven design, even while disclaiming that he’s not a DDD expert. What makes it attractive in the AI era, he says, is that DDD gives you a mature set of terms and practices the models already seem to know, so you don’t have to invent a whole language from scratch. He frames it as getting you and the AI to speak the same domain language.

The 13,000-star skill repo and the ubiquitous-language trick

The concrete example is his mattpocock/skills repo, which he says has around 13,000 stars. One standout is a “ubiquitous language” skill that scans your codebase, finds the project’s weird jargon, and turns it into a markdown glossary — whether “mole” means a skin lesion, a spy, or an animal. He keeps that document open while prompting, references it in agents.md, and says it dramatically improves clarity because the AI is using his domain model instead of guessing.

How he builds courses: Obsidian, Zettelkasten, and ruthless cutting

When asked how his teaching products come together, Pocock describes a very builder-y process: roughly two months of work for a 4.5-hour course with around 100 units, all starting as notes in an Obsidian vault using a Zettelkasten-style approach. He groups ideas, ranks lessons P1/P2/P3, records mostly the P1s, and throws out a lot. His rule of thumb is that each lesson should teach one thing, with clear dependencies and enough challenge to stretch people without overwhelming them.

Students still want lectures more than AI-native learning experiments

Pocock has a nice framework here: you can teach knowledge, skills, or wisdom. Knowledge comes through lectures, skills through exercises, and wisdom is hardest — often best developed through discussion — but he notes something funny: the more experimental and AI-heavy he makes the learning format, the more people seem to bounce off it. For now, he says, most learners still want the old formula: solid lectures, maybe homework, and clear structure.

TypeScript, agents, and the rise of “grill me”

Toward the end, the conversation widens: TypeScript may be overtaking Python in AI engineering, especially anywhere UX and shipping product matter, though Pocock carefully avoids punditry. He’s also increasingly interested in the trade-off between polished coding harnesses like Claude Code and lower-level, fully observable systems built from primitives — more convenience versus more control. And then comes the memorable prompt: his tiny “grill me” skill, which tells the AI to relentlessly interview you until you reach shared understanding; he says that pattern is useful not just for code, but for documents, planning, and all kinds of general knowledge work.